DNN convolution does not preserve XMM registers as it should.

When running a MKL_DNN operation previously created using dnnConvolutionCreateForwardBias_F32() the XMM7-8 registers (and possibly more) are not preserved.

I have code similar to this (but of course much more elaborate) which does not work in relase mode:

void RunConvolution() {
    dnnPrimitive_t handle;

    dnnConvolutionCreateForwardBias_F32(&handle, --- other pars ---);

    void* resources[dnnResourceNumber];
    --- fill in pointers to buffers, kernel and bias ---

    dnnExecute_F32(handle, resources);
}


void main()
{
    Timer timer;
    for (o =0; o < 10; o++) {
        for (int i = 0; i < 10; i++)
            RunConvolution();

        cout << timer.Elapsed() / 10.0 << endl;
    }
}

I am using the Visual Studio 2017 compiler and I noted it uses XMM7 to store the constant 10.0 _over_ the entire loops which calls RunConvolution(). I tracked the contents of the register and noted that it is destroyed by dnnExecute_F32() but that's just a thin wrapper so the actual problem is probably in some assembly code in the function redirected to when the handle refers to a structure created by dnnConvolutionCreateForwardBias_F32().

According to Microsoft documentation XMM6 and up are to be preserved over function calls, so the numbers printed are bogus. I can't see any other explanation for this than that the calling convention has been breached by the MKL DNN code.

I am using MKL 2018.1 version, 64 bit, static linking with intel threading (not TBB).

I have not tested other operations except dnnConvolutionCreateForwardBias_F32.